RePlaid Establishes Scalable Scaling Laws for Continuous Diffusion Language Models

RePlaid aligns likelihood-based continuous diffusion architectures with modern discrete DLMs to establish competitive scaling laws. The model achieves a new state-of-the-art perplexity of 22.1 on OpenWebText among continuous diffusion models and narrows the compute gap with autoregressive models to 20x.